If you have ever run a batch of files through a compression tool and wondered why some came out lean while others barely changed, this guide is for you. It is not about any particular product. It is a field guide to the decisions behind document compression — the ones that decide whether a file stays readable, stays private, and stays useful.
I will keep the jargon to a minimum, but we do need to understand what compression is actually doing. Everything else follows from that.
The only model you need
A file contains two kinds of information:
- Content — the things that carry meaning. The words, the numbers, the signature, the red seal, the fine print.
- Redundancy — repetition that can be inferred or that a human will not notice. Large areas of solid colour, neighbouring pixels of nearly the same shade, the blank margins of a scan.
Compression removes redundancy. That is the whole idea.
There are two broad flavours. Lossless compression removes only redundancy that can be perfectly reconstructed, so the file is identical when unpacked — but the size gains are limited. Lossy compression is willing to discard some information a human supposedly will not notice, buying much larger size reductions. Most everyday document compression leans lossy.
Here is the sentence that matters more than all the rest: lossy compression's judgement about what "a human won't notice" has a boundary. Understanding where that boundary sits, and how to avoid crossing it, is nearly the entire engineering problem.
Why "compress harder" eventually destroys the file
Before the boundary, the tool is cutting mostly redundancy. Text looks the same, colours look the same, the layout holds — and the file is much smaller. Everyone is happy.
Cross the boundary and the tool starts cutting real content:
- text edges pick up artefacts and soften
- thin table rules break or blur
- a scanned red seal fades toward brown
- a handwritten signature loses strokes
- fine print becomes ambiguous
None of these is a "bad compressor". They are the predictable result of asking a tool to keep removing information after there is no redundancy left to remove. The tool is still doing its job. You have simply asked for the wrong thing.
Document compression is more sensitive than photo compression for one reason: documents are dense with details that must be read exactly. A slightly soft landscape is still a nice landscape. A slightly soft contract may no longer be a reliable contract.
Practical rule: the goal is never the smallest possible file. It is the smallest file that still passes the "can a human read every part that matters" test.
Why the same setting gives different results across formats
This is the question I see most often, and it has a clean answer.
Formats encode information differently, so the amount of removable redundancy differs. A well-compressed photograph has very little left to cut; the same setting barely changes it. An untouched scan is full of redundancy; the same setting shrinks it dramatically. Nothing is broken. The inputs are simply different.
A well-built compression approach tries to make the size outcome consistent across formats at the same setting — so a PDF, an image, and a scan of similar content land in a comparable place. This is not cosmetic. It is what makes batch processing predictable. If you drop a mixed folder into a tool and one file becomes a skeleton while another does not move, you cannot plan capacity, estimate transfer time, or trust the result without inspecting every file.
How to evaluate this: take a representative batch spanning your real formats, apply one setting, and look at the distribution of size reductions. A tight distribution means predictable behaviour. A scattered one means the tool is not treating formats consistently.
Sort by destination, not by size
Here is the mental model that replaced "pick a quality slider" for me. Ask where each file is going, then choose the compression accordingly.
| Destination | What it needs | Compression posture |
|---|---|---|
| Sent now (email, chat) | Small, fast, readable on first view | More aggressive is fine |
| Kept for years (contracts, records) | Openable, verifiable, layout intact | Conservative, possibly lossless |
| Moved between devices | Compatible, no surprises | Prioritize format, not size |
| Handed to outsiders | Works on unknown systems, may be forwarded | Leave headroom |
This table is a framework, not a product recommendation. Fill it with your own file types and you will immediately see why a single global setting cannot serve all four rows.
The underlying idea is simple: different destinations have different things worth sacrificing. A file going out today can trade a little fidelity for speed. A file going into long-term storage must trade almost nothing, because you will not be there to fix it later.
The direction worth watching: purpose-aware settings
A genuinely useful direction in current tools is to stop asking the user to pick a raw "quality" number and instead let the tool infer intent from the document's purpose — conservative for archival material, aggressive for temporary transfer — while keeping the size outcome consistent across formats.
I want to be careful here: I am describing a direction and a set of criteria, not endorsing a specific implementation. The value of the idea is that it moves the decision from a number nobody understands ("quality 72") to a question everybody understands ("is this being archived or just sent?"). When you evaluate tools, check whether they support this kind of purpose-based selection, and whether their cross-format behaviour is consistent. Those two things tell you more than a headline compression ratio.
Data flow is an engineering requirement, not a feature
One more dimension gets ignored in most comparisons: where the compression happens.
Online compressors are convenient. They also require uploading the file to someone else's infrastructure. For public images, fine. For contracts, invoices, and anything carrying client information, that upload is a governance decision you may not have intended to make. You lose visibility into retention, location, and onward transfer.
The costs of getting this wrong are not hypothetical. Reporting on data protection enforcement puts the average fine tied to poor document handling in the millions of euros. In engineering terms, this is a data-flow boundary problem, and the safest boundary is the one you control — the local machine. Local compression keeps the file on the device throughout. It is worth treating "does the file leave the machine" as a first-class requirement in your comparison matrix, right alongside compression quality.
A comparison matrix you can actually use
Put these dimensions into a spreadsheet and score your candidates:
| Dimension | What to measure | Why it matters |
|---|---|---|
| Boundary control | Does it stop before content degrades? | Determines readability and trust |
| Cross-format consistency | Spread of results at one setting | Determines batch predictability |
| Data flow | Local or uploaded | Determines privacy and compliance exposure |
| Purpose-aware settings | Can it sort by archive/send/transfer? | Determines multi-scenario coverage |
| Batch stability | Throughput and failure handling | Determines large-scale feasibility |
| Archive friendliness | Layout preservation, common formats | Determines long-term readability |
Weight the rows for your context. If you mostly handle records that must survive audit, the first and last rows dominate. If you mostly move internal working files, speed and batch behaviour matter more. There is no universal winner, only a good fit for a specific workload.
Make it measurable, or it is not engineering
The last piece is the one that separates a hobby from a process.
Do not measure compression by bytes removed. Measure outcomes you can quantify:
- Clarity metrics on a fixed sample before and after — did sharpness or legibility actually drop?
- Attachment bounce rate — do recipients report unreadable files?
- Long-term openability — can an archived file still be opened correctly after a year, on a different machine?
- Batch distribution — is the size reduction consistent enough to plan around?
Once these are numbers, the question "can we compress this further?" stops being an argument and becomes a reproducible decision with a baseline and a regression check. That is the point at which compression stops being a craft and becomes something you can actually govern.
Wrapping up
Document compression is not a button. It is a set of rules about which files deserve which treatment, and a tool is only the thing that carries those rules out.
Get the model right — redundancy versus content, and the boundary between them. Sort by destination. Demand cross-format consistency. Keep sensitive files local. And measure the outcome, not the ratio.
Do those five things and the tool choice becomes almost obvious. Skip them and no tool will save you from a faded seal on a contract you needed to be readable for a decade.

Top comments (0)